How to Use AI for Scenario Planning: Best, Worst and Likely Cases

Give an AI assistant 12 months of your real figures and the four or five drivers that move your business (enquiries, win rate, average job value, material costs). Ask it to propose best, worst and likely values for each, with reasons. Then build the three cases in a spreadsheet yourself and set triggers that show which case you're in.

AI is good at the thinking part: suggesting drivers you'd missed, arguing with your assumptions, and describing how a bad quarter would actually unfold. It's unreliable at arithmetic and tends to invent "industry benchmarks", so the numbers stay in your spreadsheet. A frequent mistake is setting the worst case as a flat 20% below the likely one. Bad years are lopsided, and the question that matters isn't how far revenue falls but how many months your cash would last.

Follow me on Instagram@sagnikteaches

What a scenario is, and what it isn't

A forecast says what you think will happen. A scenario says what could happen, told as a consistent story, so you can decide in advance what you'd do. Three rules keep scenarios useful:

Connect on LinkedInSagnik Bhattacharya

• Each case is a story, not a percentage. "Worst case" is not "likely minus 20%". It's something specific: a wet winter, a large builder client going quiet, material prices jumping. The story tells you which drivers move together.
• "Likely" is not the average of best and worst. It's your honest central expectation, built from your own recent history.
• Every scenario ends in a decision. If a case doesn't change anything you'd do, you don't need it. Typical decisions: whether to hire, buy a van, raise prices, or arrange credit while trading is good.

A 12-month horizon suits most small firms. For the next few weeks of cash specifically, a 13-week cash flow forecast is the better tool; scenarios sit above it and ask bigger questions.

Subscribe on YouTube@codingliquids

Step 1: Pick the four or five drivers that matter

A driver is a number that, when it changes, moves your profit. Most service businesses have the same core set. Here's the table for a roofing contractor, with where each figure comes from:

Fixed costs (site wages you can't quickly change, overheads) sit alongside the drivers as known figures. If you're not sure you have the right drivers, this is the first job to hand to AI:

PROMPT 1: FIND THE DRIVERS
I run a [type of business] with [number] staff. Our revenue last year was
about [amount]. Our main costs are [list].
List the 5-7 factors most likely to move our profit over the next
12 months, in order of likely impact. For each, say what number I would
track to measure it and where a business like mine would normally find
that number. Don't give industry averages or percentages.

Run for a bakery with eight staff that sells over its own counter and supplies a dozen cafés, the reply might read like this (illustrative, shortened):

1. Wholesale orders per week (order sheet): the cafés are your
   largest and most predictable revenue.
2. Counter footfall and average spend (till reports).
3. Flour, butter and dairy costs as a share of sales (supplier invoices).
4. Energy cost per month (utility bills): ovens run long hours.
5. Waste as a share of production (end-of-day tally).
6. Brand awareness in the local area.

Keep the first five; each is a number the owner already has somewhere. Strike the sixth, because nothing in the business measures it and a driver you can't track can't set off a trigger. Then add the one it couldn't know: this bakery depends on two bakers who start at 4am, and a single one off sick cuts the wholesale run by a third. That becomes a capacity line in the spreadsheet, like the roofer's crews below.

Step 2: Set the ranges from your own history

For each driver, look back two or three years and find your worst quarter, your best quarter and your typical month. Those become the starting ranges. Your own history beats any figure an AI tool produces, because AI has no idea what your customers did last winter, and when asked for "typical" numbers it often produces plausible-sounding figures with no real source.

For the roofing contractor in the worked example below, the look-back might produce this (illustrative):

The notes column is what makes the table useful to AI later. It tells the model that high enquiries came with lower job values, which is exactly the "drivers move together" point a percentage range hides.

If you don't have the history, use the last 12 months and widen the ranges. A dog daycare that opened 14 months ago shows how: drop the first three months, which were a ramp-up and would drag every figure down, and look at the other 11. Its quietest month averaged 38 dogs a day and its busiest 52. Widening each end by about 10% gives a worst case of 34 and a best case of 57, but the building and staff numbers cap it at 55, so the best case stops there. One year contains only one of each season, and the wider range is an honest admission of that. If even 12 months is missing, the most useful thing you can do this quarter is start recording the numbers that feed a forecast.

Step 3: Use AI to build the stories and argue with you

With drivers and ranges in hand, use AI as a sparring partner. Remove customer names first; driver totals and ranges are enough. These prompts work in any capable assistant (ChatGPT, Claude or Gemini):

PROMPT 2: BUILD THE CASES
Here are my drivers, with my typical monthly value and the best and
worst I've seen in the last three years:
[paste driver table]
Fixed monthly costs: [amounts].
Write three scenarios for the next 12 months: best, worst and likely.
For each one:
- Tell the story in 3-4 sentences: what happens and why.
- Give a value for each driver, staying inside my ranges unless you
  explain why the story would push it outside them.
- Say which drivers would move together in this story, and why.
Do not calculate profit. I will do the arithmetic myself.

PROMPT 3: ARGUE WITH ME
Here is my "likely" case: [paste]. Act as a sceptical adviser.
What are the three assumptions most likely to be wrong, and what would
have to be true for each to hold? What constraint have I ignored
(capacity, cash timing, staff, suppliers)?

PROMPT 4: THE WORST CASE, MONTH BY MONTH
Assume the worst case starts next month. Describe months 1 to 6:
what I would notice first, when cash would start to tighten, and the
latest point at which each of these actions would still help:
[list your possible actions].

Prompt 3 is the one most people skip, and it often earns its keep. Asking AI to find the weak assumptions is the same idea as a pre-mortem: imagine it went wrong, then work out why. Treat its challenges as questions to check, not verdicts. Using ChatGPT as a business adviser without being misled covers the habit of pushing back.

Here's the kind of reply prompt 3 gives on the roofer's likely case (illustrative), with a verdict on each point:

1. Win rate held at 35% while materials rise. If you pass higher tile
   costs into quotes, win rate may fall; if you don't, margin falls.
   The likely case assumes neither happens.
2. Crew capacity of 24 jobs assumes no one leaves or is off for long.
   With three crews, losing one roofer for a month removes about a
   third of a crew's output.
3. Roofing firms typically lose 10-15% of working days to weather in
   winter, so capacity should be lower from November to February.
Constraint ignored: cash timing. Larger re-roofs are often paid on
completion, so a strong month for work can be a weak month for cash.

Points 1 and 2 are good challenges built from the owner's own inputs, and both are worth a line in the spreadsheet. The cash-timing constraint is the most valuable thing in the reply; it points straight at the 13-week forecast. Point 3 is the familiar trap: a precise-sounding "typically" figure with no source. The idea behind it is sound, so replace the number with the firm's own rained-off days from last winter's schedule.

Prompt 4 needs the same treatment. On the roofer's worst case, the month-by-month reply might open like this (illustrative):

Month 1: Enquiries fall first; quotes already out still convert,
         so revenue looks normal.
Month 2: Fewer new jobs start. Crews are fully paid but idle on
         wet days. Cash still reasonable from month 1 completions.
Month 3: Revenue drops sharply as the pipeline empties. This is
         when cash starts to tighten.
Latest useful point for each action:
- Arrange a credit line: month 2, while recent accounts look normal
- Cut crew hours: month 2 or 3
- Delay the van purchase: before any deposit is paid

The sequence is sensible, and "arrange credit while the accounts still look normal" is the most useful line in it. What it can't see are the lumpy payments only the owner knows about. This firm's annual insurance renewal and a scaffolding hire invoice both fall in month 2, which moves the cash squeeze a month earlier than the reply suggests. Add known one-off payments to the prompt, or pencil them into the spreadsheet yourself, before trusting any "cash tightens in month X" line.

Step 4: Do the maths in a spreadsheet

Put the three cases side by side, one column each, with the drivers at the top and profit calculated below. Chat assistants make arithmetic slips and rarely flag them; why AI is bad at maths explains why. If you use an assistant's data-analysis feature to run the numbers, check the totals against your own formulas.

                        LIKELY      BEST        WORST
Enquiries / month       60          70          45
Win rate                35%         38%         30%
Jobs won / month        =B2*B3      ...         ...
Crew capacity           24          24          24
Jobs done (capped)      =MIN(B4,B5)
Average job value       6,000       6,300       5,600
Revenue / month         =B6*B7
Materials % of revenue  30%         30%         32.4%
Materials               =B8*B9
Site wages              52,000      56,000      52,000
Overheads               22,000      22,000      22,000
Profit / month          =B8-B10-B11-B12

The capacity line matters. Without it, the best case assumes you can do more jobs than your crews can physically complete, which is one of the constraints AI-generated scenarios tend to miss. In Excel, the Scenario Manager (under Data, What-If Analysis) can store each case's inputs, switch between them and produce a summary report, though three plain columns are easier to read and share.

Worked example: a roofing contractor's three cases

An illustration with round figures. Say a ten-person roofing contractor has three crews, averages 60 enquiries a month, wins 35% of quotes, and averages $6,000 a job. Site wages are $52,000 a month and overheads $22,000. Using the layout above:

Three things jump out, and each leads to a decision:

• The best case is capped by capacity. Without the cap, it would have shown about $16,000 a month more revenue than the crews could deliver. Decision: line up a trusted subcontract crew now, so a surge can actually be taken.
• The break-even point is about 18 jobs a month at the likely job value and materials share ($74,000 of fixed costs divided by $4,200 of margin per job). That's only three jobs below the likely case, which is a thinner cushion than the owner assumed.
• The worst case burns cash fast. With $90,000 in reserve, a loss of about $23,000 a month lasts under four months. Even with pre-agreed cuts (moving one crew to a four-day week and trimming overheads, which brings the loss to roughly $12,000 a month), the reserve lasts about seven months. Decision: arrange an overdraft or credit line now, while the accounts look healthy, and delay the planned van purchase until two consecutive months confirm the likely case.

A fourth case worth running: the single-customer shock

Best, worst and likely cover the general climate. Many small firms also carry one specific risk that deserves its own scenario: a single customer, supplier or person the business leans on. For a roofing contractor that might be a housebuilder or letting agent sending a quarter of the work; for others it's one wholesaler, or the owner being off sick for six weeks.

Run it as a "likely case minus one": keep every driver at its likely value, then remove that customer's jobs, or that person's output, from the month it happens. The question isn't how probable it is but how long you could absorb it. If the answer is "under three months", that's a decision in its own right: widen the customer base, agree a longer notice period in the contract, or write down what the key person knows so someone else can pick it up. Ask AI to describe the first 90 days after the loss, using prompt 4 with the shock as the starting point.

The sum is short. Take a six-person commercial print shop whose largest client, a chain of estate agents, accounts for about a quarter of its work (all figures illustrative):

                           LIKELY       LIKELY MINUS ONE
Revenue / month            $48,000      $36,000
Materials (35%)            $16,800      $12,600
Fixed costs                $30,000      $30,000
Profit / month             $1,200       -$6,600

Cash reserve: $18,000  ->  lasts about 2.7 months after the loss

Under three months, so it's a decision, not a curiosity. The owner's two actions were a new-business target of three mid-sized clients over the year, and a conversation with the big client about moving to a 12-month agreement with 90 days' notice, which would turn a sudden loss into a planned one.

Step 5: Turn the cases into triggers and actions

Scenarios pay off when you decide now what you'll do later, because in the middle of a bad quarter everyone hopes next month will be better. For each case, pick an early signal, a threshold and an action:

Keep it alive with a monthly 20-minute check

At each month-end, put the actual driver values next to the three cases. Then ask AI for a short read-out, giving it only the numbers:

Here are my three scenarios and this month's actual driver values:
[paste]. In one paragraph: which scenario are we tracking closest to,
which driver is furthest from the likely case, and has any trigger
threshold been crossed? List the triggers crossed, if any, and nothing else.

Check its answer against the table yourself; it takes a minute. The reason shows up quickly. Say the roofer pastes October's actuals, with 48 enquiries, and gets back: "Tracking closest to the likely case; enquiries slightly below; no triggers crossed." It's wrong, but not because the AI slipped. September was 47, so the "under 50 for two months running" trigger has fired, and the AI never saw September. Paste the last three months of actuals every time, not just the latest one, and any trigger that depends on a run of months can be checked.

Rebuild the three cases every quarter, or immediately if something big changes, such as losing your largest customer.

Where AI scenario planning goes wrong

• Invented benchmarks. "Roofing firms typically see a 15% winter dip" sounds useful and may be made up. Use your own history.
• Symmetrical ranges. Plus or minus 20% looks tidy but reality is lopsided: a bad year can fall further than a good year can rise, especially when capacity caps the upside.
• Drivers treated as independent. In the worst case, bad weather cuts jobs and leaves you paying crews for rained-off days. Ask AI which drivers move together.
• False precision. Profit to the dollar in a scenario is noise. Round to the nearest thousand when you present it.
• Confidential figures in a personal account. Driver totals are low risk; full accounts, payroll and customer names belong only on a business plan with model training off by default, or not in AI at all.

Further reads

• How to Run a SWOT Analysis of Your Business With AI — A quicker strategic exercise to run before scenarios.
• Budget vs Actual: How to Explain Monthly Variances With AI — Explain the monthly gaps your scenario tracking will show.
• How to Write a Business Plan With AI, and What to Check Yourself — Put the likely case into a proper plan.
• What Is Cash Flow Forecasting? A Plain-English Guide With AI Examples — The cash side of each scenario, explained simply.
• How to Build a KPI Dashboard With AI When You Have No Data Team — Track your trigger signals on one screen.
• How to Build an Investor Pitch Deck With AI, and What to Check — Scenarios are the backbone of a credible pitch.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Support documentation for Excel's What-If Analysis tools. All business figures are illustrative.